小样本约束下情境化自然语言驱动的人机交互式共融机器人自动任务编程研究

Human-robot integration is a key characteristic of the next generation robot. It demands urgently for the online programming capabilities of robotic systems in order to meet the emerging requirements of performing unknown tasks that manufactures cannot completely program in advance. This project focuses on studying the fundamental theories and methods of automatic task programming of co-robot through human-robot interaction driven by situated natural language. The goal is to enable the ordinary people who do not need professional programming capability and tedious training to on site customize and extend the robot skills by only using natural language...Firstly, a computational model of information representation and inference will be studied to solve the challenging problem of effective information interaction between human, robot and environment under asymmetric information constraints. The computational model will address the challenge from a new perspective that puts an emphasis on identical representation, fusion and inference of multi-dimensional information. It would provide great support to the automatic task programming process in terms of effective information query and inference. Secondly, we will investigate a new approach to automatic task programming of co-robot through human-robot interaction, taking explicit representation and learning of abstract task knowledge as the core while considering the small sample constraint of task learning. It aims at exploring the procedure and law of how to acquire and utilize unknown task knowledge for the robot based on natural language communication between human and robot. Thirdly, the synthetic experiments of automatic task programming and execution in the real robot platforms will be performed in order to evaluate and apply the theoretical achievements...The creative works will provide great theoretical and technical value to the natural language communication based human-robot integration, and play an important role in many fields of industry, people’s livelihood and public security such as intelligent manufacturing, rehabilitation, disaster rescue and relief, etc.

与人共融是下一代机器人的本质特征，迫切需要机器人具备在线编程能力以快速响应未知任务要求。研究情境化自然语言驱动的人机交互式机器人自动任务编程理论与方法，旨在使普通用户无需专业编程技能与繁琐的训练，仅需使用自然语言即可在线快速定制和扩展机器人的工作技能。以多维度信息统一表示、融合与推理为新的视角，研究信息表示与推理计算模型，解决信息不对称条件约束下的人-机器人-环境信息交互难题，为机器人自动任务编程提供有力的信息查询与推理支持；研究探索小样本条件约束下以抽象任务显式表达与学习为核心的人机交互式机器人自动任务编程的新方法，揭示自然语言交流下机器人自动获取和运用未知任务知识的规律；在真实机器人平台上开展综合的自动任务编程与执行实验，评价与应用取得的理论成果。项目研究将为基于自然语言交流的人-机器人共融奠定理论和技术基础，在智能制造、康复医疗、救灾救援等工业、民生、公共安全领域发挥重要作用。

与人共融是下一代机器人的本质特征，迫切需要机器人具备在线编程能力以快速响应未知任务要求。研究情境化自然语言驱动的人机交互式机器人自动任务编程理论与方法，旨在使普通用户无需专业编程技能与繁琐的训练，仅需使用自然语言即可在线快速定制和扩展机器人的工作技能。研究中将与人共融机器人系统中不对称、多维度信息融合，统一表示为一阶谓词逻辑，利用马尔可夫逻辑网基础理论框架来对人-机器人-环境交互信息进行推理，为自动任务编程提供有效的关于确定或不确定、已知或未知信息的查询手段。紧扣真实应用情境下小样本条件约束，关注与任务本质属性相关的自然语言语法与语义知识、任务目标状态、约束条件、默认值、子目标序列等任务知识的归纳学习，在仿真环境中验证了从少量情境化自然语言驱动的交互实例中学习生成用于表达相应任务的混合概念任务图模型的可行性。然而，在符号层进行知识表示与推理，信息表示的准确程度会对推理结果造成较大的影响，导致相关模型只能在理想的仿真环境中或简化的真实环境中运用。因此，项目中开展了基于深度强化学习与Transformer等机器学习模型的数据驱动的机器人自动任务编程方法研究。针对物体自动分拣应用，在仿真与真实环境中均进行了实验验证，物体分拣准确率与执行效率可满足实际应用需求。如何针对特定的机器人系统与应用，充分利用知识驱动和数据驱动的优势，值得后续深入研究和探讨。本项目将共融机器人中的视觉识别、机器人运动控制、任务学习与规划等技术应用于移动机器人巡检、助老助残轮椅、工业机器人垃圾分选等真实应用，解决了全局导航规划、局部安全导航、路径跟随、重定位、行人跟随控制、物体识别与抓取等工程难题，取得了良好的社会效益和经济效益。